基于改进FCOS的水下目标检测算法  

Underwater Target Detection Algorithm Based on Improved FCOS

作　　者：陈卫东[1,2] 谢晓东 岑强 陈娜兰 朱奇光[1,2] CHEN Wei-dong;XIE Xiao-dong;CEN Qiang;CHEN Na-lan;ZHU Qi-guang(School of Information Science and Engineering,Yanshan University,Qinhuangdao,Hebei 066004,China;Key Laboratory for Special Fiber and Fiber Sensor of Hebei Province,Yanshan University,Qinhuangdao,Hebei 066004,China)

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004 [2]燕山大学河北省特种光纤与光纤传感重点实验室,河北秦皇岛066004

出　　处：《计量学报》2023年第11期1659-1666,共8页Acta Metrologica Sinica

基　　金：国家自然科学基金(61773333,62273296)。

摘　　要：提出一种基于改进全卷积单阶段目标检测(FCOS)算法的水下目标检测算法。针对水下光学图像存在高色偏、低对比度、色彩偏暗、模糊失真而导致现有目标检测算法在水下环境检测效果不佳等问题,将骨干网络中的普通卷积替换为可变形卷积(DCN)进行优化,增强算法在模糊的水下光学图片的特征提取能力。通过神经架构搜索(NAS)改进网络的特征融合网络以及检测网络,提升对骨干网络提取到的特征的利用能力。采用CIoU Loss作为新的损失函数来提高坐标回归的准确率。实验表明:改进的FCOS算法在DUO数据集上,检测的准确率提高了1.8%,召回率提高了2.2%,检测速度为53.4帧/s(相比改进前降低了5.0%)。该算法准确率较高并基本达到实时检测的要求。Propose an underwater target detection algorithm based on an improved fully convolutional one-stage object detection(FCOS).In response to problems such as high color cast,low contrast,dark color bias,and blurry distortion in underwater optical images that result in poor performance of existing target detection algorithms in underwater environments,the ordinary convolutions in the backbone network are replaced with deformable convolutions(DCN)for optimization,enhancing the algorithm's feature extraction capability in blurry underwater images.The feature fusion network and detection network of the improved network are enhanced through neural architecture search(NAS)to improve the utilization of features extracted by the backbone network.The CIoU Loss is adopted as the new loss function to improve the accuracy of coordinate regression.Experimental results show that the algorithm improves the detection accuracy by 1.8%and recall rate by 2.2%on the DUO dataset,with a detection speed of 53.4 frames/s,which is a 5.0%reduction compared to the previous version.The algorithm achieves high accuracy and meets the requirements of real-time detection.

关 键 词：计量学 水下目标检测 改进FCOS算法 DCN模块 NAS模块 

分 类 号：TB96[机械工程—光学工程] TB973[一般工业技术—计量学]